def test_save_and_load(self):
capacity = self.capacity
num_steps = self.num_steps
tempdir = tempfile.mkdtemp()
rbuf = replay_buffers.ReplayBuffer(capacity, num_steps)
correct_item = collections.deque([], maxlen=num_steps)
# Add two transitions
for _ in range(num_steps):
trans1 = dict(
state=0,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
correct_item.append(trans1)
rbuf.append(**trans1)
correct_item2 = copy.deepcopy(correct_item)
trans2 = dict(
state=1,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
correct_item2.append(trans2)
rbuf.append(**trans2)
# Now it has two transitions
assert len(rbuf) == 2
# Save
filename = os.path.join(tempdir, "rbuf.pkl")
rbuf.save(filename)
# Initialize rbuf
rbuf = replay_buffers.ReplayBuffer(capacity)
# Of course it has no transition yet
assert len(rbuf) == 0
# Load the previously saved buffer
rbuf.load(filename)
# Now it has two transitions again
assert len(rbuf) == 2
# And sampled transitions are exactly what I added!
s2 = rbuf.sample(2)
if s2[0][num_steps - 1]["state"] == 0:
assert s2[0] == list(correct_item)
assert s2[1] == list(correct_item2)
else:
assert s2[0] == list(correct_item2)
assert s2[1] == list(correct_item)
def test(self):
n = 5
if self.replay_buffer_type == "ReplayBuffer":
rbuf = replay_buffers.ReplayBuffer(capacity=None, num_steps=n)
elif self.replay_buffer_type == "PrioritizedReplayBuffer":
rbuf = replay_buffers.PrioritizedReplayBuffer(capacity=None,
num_steps=n)
else:
assert False
# 2 transitions for env_id=0
for _ in range(2):
trans1 = dict(
state=0,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
rbuf.append(env_id=0, **trans1)
# 4 transitions for env_id=1 with a terminal state
for i in range(4):
trans1 = dict(
state=0,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=(i == 3),
)
rbuf.append(env_id=1, **trans1)
# 9 transitions for env_id=2
for _ in range(9):
trans1 = dict(
state=0,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
rbuf.append(env_id=2, **trans1)
# It should have:
# - 4 transitions from env_id=1
# - 5 transitions from env_id=2
assert len(rbuf) == 9
# env_id=0 episode ends
rbuf.stop_current_episode(env_id=0)
# Now it should have 9 + 2 = 11 transitions
assert len(rbuf) == 11
# env_id=2 episode ends
rbuf.stop_current_episode(env_id=2)
# Finally it should have 9 + 2 + 4 = 15 transitions
assert len(rbuf) == 15
def _test_load_rainbow(self, gpu):
from pfrl.q_functions import DistributionalDuelingDQN
q_func = DistributionalDuelingDQN(4, 51, -10, 10)
pnn.to_factorized_noisy(q_func, sigma_scale=0.5)
explorer = explorers.Greedy()
opt = torch.optim.Adam(q_func.parameters(), 6.25e-5, eps=1.5 * 10**-4)
rbuf = replay_buffers.ReplayBuffer(100)
agent = agents.CategoricalDoubleDQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=0.99,
explorer=explorer,
minibatch_size=32,
replay_start_size=50,
target_update_interval=32000,
update_interval=4,
batch_accumulator="mean",
phi=lambda x: x,
)
downloaded_model, exists = download_model(
"Rainbow",
"BreakoutNoFrameskip-v4",
model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_ddpg(self, gpu):
obs_size = 11
action_size = 3
from pfrl.nn import ConcatObsAndAction
q_func = nn.Sequential(
ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, 1),
)
from pfrl.nn import BoundByTanh
from pfrl.policies import DeterministicHead
policy = nn.Sequential(
nn.Linear(obs_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, action_size),
BoundByTanh(low=[-1.0, -1.0, -1.0], high=[1.0, 1.0, 1.0]),
DeterministicHead(),
)
opt_a = torch.optim.Adam(policy.parameters())
opt_c = torch.optim.Adam(q_func.parameters())
explorer = explorers.AdditiveGaussian(scale=0.1,
low=[-1.0, -1.0, -1.0],
high=[1.0, 1.0, 1.0])
agent = agents.DDPG(
policy,
q_func,
opt_a,
opt_c,
replay_buffers.ReplayBuffer(100),
gamma=0.99,
explorer=explorer,
replay_start_size=1000,
target_update_method="soft",
target_update_interval=1,
update_interval=1,
soft_update_tau=5e-3,
n_times_update=1,
gpu=gpu,
minibatch_size=100,
burnin_action_func=None,
)
downloaded_model, exists = download_model(
"DDPG", "Hopper-v2", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_td3(self, gpu):
obs_size = 11
action_size = 3
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pnn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, 1),
)
q_func_optimizer = torch.optim.Adam(q_func.parameters())
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
policy = nn.Sequential(
nn.Linear(obs_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, action_size),
nn.Tanh(),
pfrl.policies.DeterministicHead(),
)
policy_optimizer = torch.optim.Adam(policy.parameters())
rbuf = replay_buffers.ReplayBuffer(100)
explorer = explorers.AdditiveGaussian(scale=0.1,
low=[-1.0, -1.0, -1.0],
high=[1.0, 1.0, 1.0])
agent = agents.TD3(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=0.99,
soft_update_tau=5e-3,
explorer=explorer,
replay_start_size=1000,
gpu=gpu,
minibatch_size=100,
burnin_action_func=None,
)
downloaded_model, exists = download_model(
"TD3", "Hopper-v2", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def _test_load_iqn(self, gpu):
n_actions = 4
q_func = pfrl.agents.iqn.ImplicitQuantileQFunction(
psi=nn.Sequential(
nn.Conv2d(4, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.ReLU(),
nn.Flatten(),
),
phi=nn.Sequential(
pfrl.agents.iqn.CosineBasisLinear(64, 3136),
nn.ReLU(),
),
f=nn.Sequential(
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
),
)
# Use the same hyper parameters as https://arxiv.org/abs/1710.10044
opt = torch.optim.Adam(q_func.parameters(), lr=5e-5, eps=1e-2 / 32)
rbuf = replay_buffers.ReplayBuffer(100)
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=0.1,
decay_steps=10**6,
random_action_func=lambda: np.random.randint(4),
)
agent = agents.IQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=50,
target_update_interval=10**4,
update_interval=4,
batch_accumulator="mean",
phi=lambda x: x,
quantile_thresholds_N=64,
quantile_thresholds_N_prime=64,
quantile_thresholds_K=32,
)
downloaded_model, exists = download_model(
"IQN", "BreakoutNoFrameskip-v4", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def test_append_and_terminate(self):
capacity = self.capacity
num_steps = self.num_steps
rbuf = replay_buffers.ReplayBuffer(capacity, num_steps)
assert len(rbuf) == 0
# Add one and sample one
for _ in range(num_steps):
trans1 = dict(
state=0,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
rbuf.append(**trans1)
assert len(rbuf) == 1
s1 = rbuf.sample(1)
assert len(s1) == 1
# Add two and sample two, which must be unique
trans2 = dict(
state=1,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=True,
)
rbuf.append(**trans2)
assert len(rbuf) == self.num_steps + 1
s2 = rbuf.sample(self.num_steps + 1)
assert len(s2) == self.num_steps + 1
if self.num_steps == 1:
if s2[0][0]["state"] == 0:
assert s2[1][0]["state"] == 1
else:
assert s2[1][0]["state"] == 0
else:
for item in s2:
# e.g. if states are 0,0,0,1 then buffer looks like:
# [[0,0,0], [0, 0, 1], [0, 1], [1]]
if len(item) < self.num_steps:
assert item[len(item) - 1]["state"] == 1
for i in range(len(item) - 1):
assert item[i]["state"] == 0
else:
for i in range(len(item) - 1):
assert item[i]["state"] == 0
def _test_load_dqn(self, gpu):
from pfrl.q_functions import DiscreteActionValueHead
n_actions = 4
q_func = nn.Sequential(
pnn.LargeAtariCNN(),
init_chainer_default(nn.Linear(512, n_actions)),
DiscreteActionValueHead(),
)
# Use the same hyperparameters as the Nature paper
opt = pfrl.optimizers.RMSpropEpsInsideSqrt(
q_func.parameters(),
lr=2.5e-4,
alpha=0.95,
momentum=0.0,
eps=1e-2,
centered=True,
)
rbuf = replay_buffers.ReplayBuffer(100)
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=0.1,
decay_steps=10**6,
random_action_func=lambda: np.random.randint(4),
)
agent = agents.DQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=50,
target_update_interval=10**4,
clip_delta=True,
update_interval=4,
batch_accumulator="sum",
phi=lambda x: x,
)
downloaded_model, exists = download_model(
"DQN", "BreakoutNoFrameskip-v4", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def test_append_and_sample(self):
capacity = self.capacity
num_steps = self.num_steps
rbuf = replay_buffers.ReplayBuffer(capacity, num_steps)
assert len(rbuf) == 0
# Add one and sample one
correct_item = collections.deque([], maxlen=num_steps)
for _ in range(num_steps):
trans1 = dict(
state=0,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
correct_item.append(trans1)
rbuf.append(**trans1)
assert len(rbuf) == 1
s1 = rbuf.sample(1)
assert len(s1) == 1
assert s1[0] == list(correct_item)
# Add two and sample two, which must be unique
correct_item2 = copy.deepcopy(correct_item)
trans2 = dict(
state=1,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
correct_item2.append(trans2)
rbuf.append(**trans2)
assert len(rbuf) == 2
s2 = rbuf.sample(2)
assert len(s2) == 2
if s2[0][num_steps - 1]["state"] == 0:
assert s2[0] == list(correct_item)
assert s2[1] == list(correct_item2)
else:
assert s2[1] == list(correct_item)
assert s2[0] == list(correct_item2)
def test_stop_current_episode(self):
capacity = self.capacity
num_steps = self.num_steps
rbuf = replay_buffers.ReplayBuffer(capacity, num_steps)
assert len(rbuf) == 0
# Add one and sample one
for _ in range(num_steps - 1):
trans1 = dict(
state=0,
action=1,
reward=2,
next_state=3,
next_action=4,
is_state_terminal=False,
)
rbuf.append(**trans1)
# we haven't experienced n transitions yet
assert len(rbuf) == 0
# episode ends
rbuf.stop_current_episode()
# episode ends, so we should add n-1 transitions
assert len(rbuf) == self.num_steps - 1
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--final-exploration-frames", type=int, default=10**6)
parser.add_argument("--final-epsilon", type=float, default=0.01)
parser.add_argument("--eval-epsilon", type=float, default=0.001)
parser.add_argument("--noisy-net-sigma", type=float, default=None)
parser.add_argument(
"--arch",
type=str,
default="doubledqn",
choices=["nature", "nips", "dueling", "doubledqn"],
)
parser.add_argument("--steps", type=int, default=5 * 10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=5 * 10**4)
parser.add_argument("--target-update-interval",
type=int,
default=3 * 10**4)
parser.add_argument("--eval-interval", type=int, default=10**5)
parser.add_argument("--update-interval", type=int, default=4)
parser.add_argument("--eval-n-runs", type=int, default=10)
parser.add_argument("--no-clip-delta",
dest="clip_delta",
action="store_false")
parser.set_defaults(clip_delta=True)
parser.add_argument("--agent",
type=str,
default="DoubleDQN",
choices=["DQN", "DoubleDQN", "PAL"])
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--lr",
type=float,
default=2.5e-4,
help="Learning rate")
parser.add_argument(
"--prioritized",
action="store_true",
default=False,
help="Use prioritized experience replay.",
)
parser.add_argument("--num-envs", type=int, default=1)
parser.add_argument("--n-step-return", type=int, default=1)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(idx, test):
# Use different random seeds for train and test envs
process_seed = int(process_seeds[idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
frame_stack=False,
)
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
env.seed(env_seed)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
def make_batch_env(test):
vec_env = pfrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, idx, test)
for idx, env in enumerate(range(args.num_envs))
])
vec_env = pfrl.wrappers.VectorFrameStack(vec_env, 4)
return vec_env
sample_env = make_env(0, test=False)
n_actions = sample_env.action_space.n
q_func = parse_arch(args.arch, n_actions)
if args.noisy_net_sigma is not None:
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
# Use the same hyper parameters as the Nature paper's
opt = optim.RMSprop(
q_func.parameters(),
lr=args.lr,
alpha=0.95,
momentum=0.0,
eps=1e-2,
centered=True,
)
# Select a replay buffer to use
if args.prioritized:
# Anneal beta from beta0 to 1 throughout training
betasteps = args.steps / args.update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(
10**6,
alpha=0.6,
beta0=0.4,
betasteps=betasteps,
num_steps=args.n_step_return,
)
else:
rbuf = replay_buffers.ReplayBuffer(10**6, num_steps=args.n_step_return)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = parse_agent(args.agent)
agent = Agent(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
clip_delta=args.clip_delta,
update_interval=args.update_interval,
batch_accumulator="sum",
phi=phi,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(
env=make_batch_env(test=True),
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_batch_with_evaluation(
agent=agent,
env=make_batch_env(test=False),
eval_env=make_batch_env(test=True),
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
log_interval=1000,
)
def main():
import logging
logging.basicConfig(level=logging.INFO)
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--env", type=str, default="CartPole-v1")
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--final-exploration-steps", type=int, default=1000)
parser.add_argument("--start-epsilon", type=float, default=1.0)
parser.add_argument("--end-epsilon", type=float, default=0.1)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--steps", type=int, default=10**8)
parser.add_argument("--prioritized-replay", action="store_true")
parser.add_argument("--replay-start-size", type=int, default=50)
parser.add_argument("--target-update-interval", type=int, default=100)
parser.add_argument("--target-update-method", type=str, default="hard")
parser.add_argument("--soft-update-tau", type=float, default=1e-2)
parser.add_argument("--update-interval", type=int, default=1)
parser.add_argument("--eval-n-runs", type=int, default=100)
parser.add_argument("--eval-interval", type=int, default=1000)
parser.add_argument("--n-hidden-channels", type=int, default=12)
parser.add_argument("--n-hidden-layers", type=int, default=3)
parser.add_argument("--gamma", type=float, default=0.95)
parser.add_argument("--minibatch-size", type=int, default=None)
parser.add_argument("--render-train", action="store_true")
parser.add_argument("--render-eval", action="store_true")
parser.add_argument("--monitor", action="store_true")
parser.add_argument("--reward-scale-factor", type=float, default=1.0)
args = parser.parse_args()
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
env = gym.make(args.env)
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = pfrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if (args.render_eval and test) or (args.render_train and not test):
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.max_episode_steps
obs_size = env.observation_space.low.size
action_space = env.action_space
n_atoms = 51
v_max = 500
v_min = 0
n_actions = action_space.n
q_func = q_functions.DistributionalFCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_atoms,
v_min,
v_max,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon,
args.end_epsilon,
args.final_exploration_steps,
action_space.sample,
)
opt = torch.optim.Adam(q_func.parameters(), 1e-3)
rbuf_capacity = 50000 # 5 * 10 ** 5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps -
args.replay_start_size) // args.update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffers.ReplayBuffer(rbuf_capacity)
agent = pfrl.agents.CategoricalDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
if args.load:
agent.load(args.load)
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
train_max_episode_len=timestep_limit,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--final-exploration-frames", type=int, default=10**6)
parser.add_argument("--final-epsilon", type=float, default=0.1)
parser.add_argument("--eval-epsilon", type=float, default=0.05)
parser.add_argument("--steps", type=int, default=10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=5 * 10**4)
parser.add_argument("--target-update-interval", type=int, default=10**4)
parser.add_argument("--eval-interval", type=int, default=10**5)
parser.add_argument("--update-interval", type=int, default=4)
parser.add_argument("--eval-n-runs", type=int, default=10)
parser.add_argument("--batch-size", type=int, default=32)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
n_atoms = 51
v_max = 10
v_min = -10
q_func = torch.nn.Sequential(
pfrl.nn.LargeAtariCNN(),
pfrl.q_functions.DistributionalFCStateQFunctionWithDiscreteAction(
512,
n_actions,
n_atoms,
v_min,
v_max,
n_hidden_channels=0,
n_hidden_layers=0,
),
)
# Use the same hyper parameters as https://arxiv.org/abs/1707.06887
opt = torch.optim.Adam(q_func.parameters(),
2.5e-4,
eps=1e-2 / args.batch_size)
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.CategoricalDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator="mean",
phi=phi,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
eval_env=eval_env,
)
def main():
import logging
logging.basicConfig(level=logging.INFO)
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--env", type=str, default="Pendulum-v0")
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--final-exploration-steps", type=int, default=10**4)
parser.add_argument("--start-epsilon", type=float, default=1.0)
parser.add_argument("--end-epsilon", type=float, default=0.1)
parser.add_argument("--noisy-net-sigma", type=float, default=None)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--steps", type=int, default=10**5)
parser.add_argument("--prioritized-replay", action="store_true")
parser.add_argument("--replay-start-size", type=int, default=1000)
parser.add_argument("--target-update-interval", type=int, default=10**2)
parser.add_argument("--target-update-method", type=str, default="hard")
parser.add_argument("--soft-update-tau", type=float, default=1e-2)
parser.add_argument("--update-interval", type=int, default=1)
parser.add_argument("--eval-n-runs", type=int, default=100)
parser.add_argument("--eval-interval", type=int, default=10**4)
parser.add_argument("--n-hidden-channels", type=int, default=100)
parser.add_argument("--n-hidden-layers", type=int, default=2)
parser.add_argument("--gamma", type=float, default=0.99)
parser.add_argument("--minibatch-size", type=int, default=None)
parser.add_argument("--render-train", action="store_true")
parser.add_argument("--render-eval", action="store_true")
parser.add_argument("--monitor", action="store_true")
parser.add_argument("--reward-scale-factor", type=float, default=1e-3)
parser.add_argument(
"--actor-learner",
action="store_true",
help="Enable asynchronous sampling with asynchronous actor(s)",
) # NOQA
parser.add_argument(
"--num-envs",
type=int,
default=1,
help=("The number of environments for sampling (only effective with"
" --actor-learner enabled)"),
) # NOQA
args = parser.parse_args()
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def clip_action_filter(a):
return np.clip(a, action_space.low, action_space.high)
def make_env(idx=0, test=False):
env = gym.make(args.env)
# Use different random seeds for train and test envs
process_seed = int(process_seeds[idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
utils.set_random_seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if isinstance(env.action_space, spaces.Box):
utils.env_modifiers.make_action_filtered(env, clip_action_filter)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = pfrl.wrappers.ScaleReward(env, args.reward_scale_factor)
if (args.render_eval and test) or (args.render_train and not test):
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.max_episode_steps
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
if isinstance(action_space, spaces.Box):
action_size = action_space.low.size
# Use NAF to apply DQN to continuous action spaces
q_func = q_functions.FCQuadraticStateQFunction(
obs_size,
action_size,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
action_space=action_space,
)
# Use the Ornstein-Uhlenbeck process for exploration
ou_sigma = (action_space.high - action_space.low) * 0.2
explorer = explorers.AdditiveOU(sigma=ou_sigma)
else:
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size,
n_actions,
n_hidden_channels=args.n_hidden_channels,
n_hidden_layers=args.n_hidden_layers,
)
# Use epsilon-greedy for exploration
explorer = explorers.LinearDecayEpsilonGreedy(
args.start_epsilon,
args.end_epsilon,
args.final_exploration_steps,
action_space.sample,
)
if args.noisy_net_sigma is not None:
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
opt = optim.Adam(q_func.parameters())
rbuf_capacity = 5 * 10**5
if args.minibatch_size is None:
args.minibatch_size = 32
if args.prioritized_replay:
betasteps = (args.steps -
args.replay_start_size) // args.update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(rbuf_capacity,
betasteps=betasteps)
else:
rbuf = replay_buffers.ReplayBuffer(rbuf_capacity)
agent = DQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
minibatch_size=args.minibatch_size,
target_update_method=args.target_update_method,
soft_update_tau=args.soft_update_tau,
)
if args.load:
agent.load(args.load)
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
elif not args.actor_learner:
print(
"WARNING: Since https://github.com/pfnet/pfrl/pull/112 we have started"
" setting `eval_during_episode=True` in this script, which affects the"
" timings of evaluation phases.")
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
train_max_episode_len=timestep_limit,
eval_during_episode=True,
)
else:
# using impala mode when given num of envs
# When we use multiple envs, it is critical to ensure each env
# can occupy a CPU core to get the best performance.
# Therefore, we need to prevent potential CPU over-provision caused by
# multi-threading in Openmp and Numpy.
# Disable the multi-threading on Openmp and Numpy.
os.environ["OMP_NUM_THREADS"] = "1" # NOQA
(
make_actor,
learner,
poller,
exception_event,
) = agent.setup_actor_learner_training(args.num_envs)
poller.start()
learner.start()
experiments.train_agent_async(
processes=args.num_envs,
make_agent=make_actor,
make_env=make_env,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
stop_event=learner.stop_event,
exception_event=exception_event,
)
poller.stop()
learner.stop()
poller.join()
learner.join()
def _objective_core(
# optuna parameters
trial,
# training parameters
env_id,
outdir,
seed,
monitor,
gpu,
steps,
train_max_episode_len,
eval_n_episodes,
eval_interval,
batch_size,
# hyperparameters
hyperparams,
):
# Set a random seed used in PFRL
utils.set_random_seed(seed)
# Set different random seeds for train and test envs.
train_seed = seed
test_seed = 2**31 - 1 - seed
def make_env(test=False):
env = gym.make(env_id)
if not isinstance(env.observation_space, gym.spaces.Box):
raise ValueError(
"Supported only Box observation environments, but given: {}".format(
env.observation_space
)
)
if len(env.observation_space.shape) != 1:
raise ValueError(
"Supported only observation spaces with ndim==1, but given: {}".format(
env.observation_space.shape
)
)
if not isinstance(env.action_space, gym.spaces.Discrete):
raise ValueError(
"Supported only discrete action environments, but given: {}".format(
env.action_space
)
)
env_seed = test_seed if test else train_seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if monitor:
env = pfrl.wrappers.Monitor(env, outdir)
if not test:
# Scale rewards (and thus returns) to a reasonable range so that
# training is easier
env = pfrl.wrappers.ScaleReward(env, hyperparams["reward_scale_factor"])
return env
env = make_env(test=False)
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
n_actions = action_space.n
# create model & q_function
model = MLP(
in_size=obs_size, out_size=n_actions, hidden_sizes=hyperparams["hidden_sizes"]
)
q_func = q_functions.SingleModelStateQFunctionWithDiscreteAction(model=model)
# Use epsilon-greedy for exploration
start_epsilon = 1
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=start_epsilon,
end_epsilon=hyperparams["end_epsilon"],
decay_steps=hyperparams["decay_steps"],
random_action_func=action_space.sample,
)
opt = optim.Adam(
q_func.parameters(), lr=hyperparams["lr"], eps=hyperparams["adam_eps"]
)
rbuf_capacity = steps
rbuf = replay_buffers.ReplayBuffer(rbuf_capacity)
agent = DQN(
q_func,
opt,
rbuf,
gpu=gpu,
gamma=hyperparams["gamma"],
explorer=explorer,
replay_start_size=hyperparams["replay_start_size"],
target_update_interval=hyperparams["target_update_interval"],
update_interval=hyperparams["update_interval"],
minibatch_size=batch_size,
)
eval_env = make_env(test=True)
evaluation_hooks = [OptunaPrunerHook(trial=trial)]
_, eval_stats_history = experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=steps,
eval_n_steps=None,
eval_n_episodes=eval_n_episodes,
eval_interval=eval_interval,
outdir=outdir,
eval_env=eval_env,
train_max_episode_len=train_max_episode_len,
evaluation_hooks=evaluation_hooks,
)
score = _get_score_from_eval_stats_history(eval_stats_history)
return score
def make_replay_buffer(self, env):
return replay_buffers.ReplayBuffer(10**5, num_steps=3)
def make_replay_buffer(self, env):
return replay_buffers.ReplayBuffer(10**5)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument(
"--env",
type=str,
default="Hopper-v2",
help="OpenAI Gym MuJoCo env to perform algorithm on.",
)
parser.add_argument("--num-envs",
type=int,
default=1,
help="Number of envs run in parallel.")
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--load",
type=str,
default="",
help="Directory to load agent from.")
parser.add_argument(
"--steps",
type=int,
default=10**6,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=10,
help="Number of episodes run for each evaluation.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=5000,
help="Interval in timesteps between evaluations.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=10000,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument("--batch-size",
type=int,
default=256,
help="Minibatch size")
parser.add_argument("--render",
action="store_true",
help="Render env states in a GUI window.")
parser.add_argument("--demo",
action="store_true",
help="Just run evaluation, not training.")
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--monitor",
action="store_true",
help="Wrap env with gym.wrappers.Monitor.")
parser.add_argument(
"--log-interval",
type=int,
default=1000,
help=
"Interval in timesteps between outputting log messages during training",
)
parser.add_argument("--log-level",
type=int,
default=logging.INFO,
help="Level of the root logger.")
parser.add_argument(
"--policy-output-scale",
type=float,
default=1.0,
help="Weight initialization scale of policy output.",
)
parser.add_argument(
"--optimizer",
type=str,
default="AdaBelief",
)
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def make_env(process_idx, test):
env = gym.make(args.env)
# Unwrap TimiLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.env
# Use different random seeds for train and test envs
process_seed = int(process_seeds[process_idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
# Normalize action space to [-1, 1]^n
env = pfrl.wrappers.NormalizeActionSpace(env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if args.render:
env = pfrl.wrappers.Render(env)
return env
def make_batch_env(test):
return pfrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, idx, test)
for idx, env in enumerate(range(args.num_envs))
])
sample_env = make_env(process_idx=0, test=False)
timestep_limit = sample_env.spec.max_episode_steps
obs_space = sample_env.observation_space
action_space = sample_env.action_space
print("Observation space:", obs_space)
print("Action space:", action_space)
obs_size = obs_space.low.size
action_size = action_space.low.size
if LooseVersion(torch.__version__) < LooseVersion("1.5.0"):
raise Exception("This script requires a PyTorch version >= 1.5.0")
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = torch.chunk(x, 2, dim=1)
log_scale = torch.clamp(log_scale, -20.0, 2.0)
var = torch.exp(log_scale * 2)
base_distribution = distributions.Independent(
distributions.Normal(loc=mean, scale=torch.sqrt(var)), 1)
# cache_size=1 is required for numerical stability
return distributions.transformed_distribution.TransformedDistribution(
base_distribution,
[distributions.transforms.TanhTransform(cache_size=1)])
def make_optimizer(parameters):
if args.optimizer == "OfficialAdaBelief":
import adabelief_pytorch
optim_class = adabelief_pytorch.AdaBelief
optim = optim_class(parameters, betas=(0.9, 0.999), eps=1e-12)
else:
optim_class = getattr(
torch_optimizer,
args.optimizer,
getattr(torch.optim, args.optimizer, None),
)
optim = optim_class(parameters)
assert optim_class is not None
print(str(optim_class), "with default hyperparameters")
return optim
policy = nn.Sequential(
nn.Linear(obs_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, action_size * 2),
Lambda(squashed_diagonal_gaussian_head),
)
torch.nn.init.xavier_uniform_(policy[0].weight)
torch.nn.init.xavier_uniform_(policy[2].weight)
torch.nn.init.xavier_uniform_(policy[4].weight,
gain=args.policy_output_scale)
policy_optimizer = make_optimizer(policy.parameters())
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 1),
)
torch.nn.init.xavier_uniform_(q_func[1].weight)
torch.nn.init.xavier_uniform_(q_func[3].weight)
torch.nn.init.xavier_uniform_(q_func[5].weight)
q_func_optimizer = make_optimizer(q_func.parameters())
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
rbuf = replay_buffers.ReplayBuffer(10**6)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
# Hyperparameters in http://arxiv.org/abs/1802.09477
agent = pfrl.agents.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=0.99,
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=burnin_action_func,
entropy_target=-action_size,
temperature_optimizer_lr=3e-4,
)
if len(args.load) > 0 or args.load_pretrained:
if args.load_pretrained:
raise Exception("Pretrained models are currently unsupported.")
# either load or load_pretrained must be false
assert not len(args.load) > 0 or not args.load_pretrained
if len(args.load) > 0:
agent.load(args.load)
else:
agent.load(
utils.download_model("SAC",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(
env=make_batch_env(test=True),
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_batch_with_evaluation(
agent=agent,
env=make_batch_env(test=False),
eval_env=make_batch_env(test=True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
log_interval=args.log_interval,
max_episode_len=timestep_limit,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument(
"--env",
type=str,
default="RoboschoolAtlasForwardWalk-v1",
help="OpenAI Gym env to perform algorithm on.",
)
parser.add_argument("--num-envs",
type=int,
default=4,
help="Number of envs run in parallel.")
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--load",
type=str,
default="",
help="Directory to load agent from.")
parser.add_argument(
"--steps",
type=int,
default=10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=20,
help="Number of episodes run for each evaluation.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=100000,
help="Interval in timesteps between evaluations.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=10000,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument(
"--update-interval",
type=int,
default=1,
help="Interval in timesteps between model updates.",
)
parser.add_argument("--batch-size",
type=int,
default=256,
help="Minibatch size")
parser.add_argument("--render",
action="store_true",
help="Render env states in a GUI window.")
parser.add_argument("--demo",
action="store_true",
help="Just run evaluation, not training.")
parser.add_argument("--monitor",
action="store_true",
help="Wrap env with Monitor to write videos.")
parser.add_argument(
"--log-interval",
type=int,
default=1000,
help=
"Interval in timesteps between outputting log messages during training",
)
parser.add_argument("--log-level",
type=int,
default=logging.INFO,
help="Level of the root logger.")
parser.add_argument(
"--n-hidden-channels",
type=int,
default=1024,
help="Number of hidden channels of NN models.",
)
parser.add_argument("--discount",
type=float,
default=0.98,
help="Discount factor.")
parser.add_argument("--n-step-return",
type=int,
default=3,
help="N-step return.")
parser.add_argument("--lr",
type=float,
default=3e-4,
help="Learning rate.")
parser.add_argument("--adam-eps",
type=float,
default=1e-1,
help="Adam eps.")
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def make_batch_env(test):
return pfrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, args, process_seeds[idx], test)
for idx, env in enumerate(range(args.num_envs))
])
sample_env = make_env(args, process_seeds[0], test=False)
timestep_limit = sample_env.spec.max_episode_steps
obs_space = sample_env.observation_space
action_space = sample_env.action_space
print("Observation space:", obs_space)
print("Action space:", action_space)
del sample_env
action_size = action_space.low.size
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = torch.chunk(x, 2, dim=1)
log_scale = torch.clamp(log_scale, -20.0, 2.0)
var = torch.exp(log_scale * 2)
base_distribution = distributions.Independent(
distributions.Normal(loc=mean, scale=torch.sqrt(var)), 1)
# cache_size=1 is required for numerical stability
return distributions.transformed_distribution.TransformedDistribution(
base_distribution,
[distributions.transforms.TanhTransform(cache_size=1)])
policy = nn.Sequential(
nn.Linear(obs_space.low.size, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, action_size * 2),
Lambda(squashed_diagonal_gaussian_head),
)
torch.nn.init.xavier_uniform_(policy[0].weight)
torch.nn.init.xavier_uniform_(policy[2].weight)
torch.nn.init.xavier_uniform_(policy[4].weight)
policy_optimizer = torch.optim.Adam(policy.parameters(),
lr=args.lr,
eps=args.adam_eps)
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_space.low.size + action_size,
args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, 1),
)
torch.nn.init.xavier_uniform_(q_func[1].weight)
torch.nn.init.xavier_uniform_(q_func[3].weight)
torch.nn.init.xavier_uniform_(q_func[5].weight)
q_func_optimizer = torch.optim.Adam(q_func.parameters(),
lr=args.lr,
eps=args.adam_eps)
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
rbuf = replay_buffers.ReplayBuffer(10**6, num_steps=args.n_step_return)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
# Hyperparameters in http://arxiv.org/abs/1802.09477
agent = pfrl.agents.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=args.discount,
update_interval=args.update_interval,
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=burnin_action_func,
entropy_target=-action_size,
temperature_optimizer_lr=args.lr,
)
if len(args.load) > 0:
agent.load(args.load)
if args.demo:
eval_env = make_env(args, seed=0, test=True)
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_batch_with_evaluation(
agent=agent,
env=make_batch_env(test=False),
eval_env=make_batch_env(test=True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
log_interval=args.log_interval,
max_episode_len=timestep_limit,
)
def __init__(self, buffer_size):
self._idx = 0 # 次にデータを挿入するインデックス．
self._size = 0 # データ数．
self.buffer_size = buffer_size # リプレイバッファのサイズ．
self.buf = replay_buffers.ReplayBuffer(capacity=self.buffer_size)
self.dev = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=(
"Directory path to save output files."
" If it does not exist, it will be created."
),
)
parser.add_argument(
"--env",
type=str,
default="'DClawTurnFixed-v0'",
help="OpenAI Gym MuJoCo env to perform algorithm on.",
)
parser.add_argument("--seed", type=int, default=0, help="Random seed [0, 2 ** 32)")
parser.add_argument(
"--gpu", type=int, default=-1, help="GPU to use, set to -1 if no GPU."
)
parser.add_argument(
"--load", type=str, default="", help="Directory to load agent from."
)
parser.add_argument(
"--max-steps",
type=int,
default=10 ** 6,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=10,
help="Number of episodes run for each evaluation.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=5000,
help="Interval in timesteps between evaluations.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=10000,
help="Minimum replay buffer size before " + "performing gradient updates.",
)
parser.add_argument("--batch-size", type=int, default=64, help="Minibatch size")
parser.add_argument(
"--render", action="store_true", help="Render env states in a GUI window."
)
parser.add_argument(
"--demo", action="store_true", help="Just run evaluation, not training."
)
parser.add_argument("--load-pretrained", action="store_true", default=False)
parser.add_argument(
"--pretrained-type", type=str, default="best", choices=["best", "final"]
)
parser.add_argument(
"--monitor", action="store_true", help="Wrap env with gym.wrappers.Monitor."
)
parser.add_argument(
"--log-level", type=int, default=logging.INFO, help="Level of the root logger."
)
parser.add_argument("--gamma", type=float, default=0.9)
parser.add_argument("--ddpg-training-steps", type=int, default=int(1e3))
parser.add_argument("--adversary-training-steps", type=int,default=int(1e3))
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
args.outdir = './results'
print("Output files are saved in {}".format(args.outdir))
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
def make_env(test):
env = gym.make('DClawTurnFixed-v0')
# Unwrap TimeLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.env
# Use different random seeds for train and test envs
env_seed = 2 ** 32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if args.render and not test:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.max_episode_steps
obs_space = env.observation_space
action_space = env.action_space
print("Observation space:", obs_space)
print("Action space:", action_space)
obs_size = obs_space.low.size
action_size = action_space.low.size
q_func = nn.Sequential(
ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256,256),
nn.ReLU(),
nn.Linear(256, 1),
)
policy = nn.Sequential(
nn.Linear(obs_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256,256),
nn.ReLU(),
nn.Linear(256, action_size),
BoundByTanh(low=action_space.low, high=action_space.high),
DeterministicHead(),
)
ddpg_opt_a = torch.optim.Adam(policy.parameters())
ddpg_opt_c = torch.optim.Adam(q_func.parameters())
ddpg_rbuf = replay_buffers.ReplayBuffer(10 ** 6)
ddpg_explorer = explorers.AdditiveGaussian(
scale=0.1, low=action_space.low, high=action_space.high
)
def ddpg_burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low, action_space.high).astype(np.float32)
# Hyperparameters in http://arxiv.org/abs/1802.09477
ddpg_agent = DDPG(
policy,
q_func,
ddpg_opt_a,
ddpg_opt_c,
ddpg_rbuf,
gamma=args.gamma,
explorer=ddpg_explorer,
replay_start_size=args.replay_start_size,
target_update_method="soft",
target_update_interval=1,
update_interval=1,
soft_update_tau=5e-3,
n_times_update=1,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=ddpg_burnin_action_func,
)
def adversary_random_func():
return np.random.randint(0,9)
# adversary_q = Critic(obs_size, 1, hidden_size=adversary_hidden_size)
# adversary_action_space = gym.spaces.discrete.Discrete(9)
# adversary_q = q_functions.FCQuadraticStateQFunction(
# obs_size, 1, n_hidden_channels = 256, n_hidden_layers = 2,action_space = adversary_action_space
# )
adversary_q = nn.Sequential(
nn.Linear(obs_size, 256),
nn.Linear(256,256),
nn.Linear(256,256),
nn.Linear(256,1),
DiscreteActionValueHead(),
)
adversary_optimizer = torch.optim.Adam(adversary_q.parameters(), lr=1e-3)
adversary_rbuf_capacity = int(1e6)
adversary_rbuf = replay_buffers.ReplayBuffer(adversary_rbuf_capacity)
adversary_explorer = explorers.LinearDecayEpsilonGreedy(
1.0, 0.1, 10**4, adversary_random_func
)
adversary_agent = DQN(
adversary_q,
adversary_optimizer,
adversary_rbuf,
gpu=args.gpu,
gamma=args.gamma,
explorer=adversary_explorer,
replay_start_size=args.replay_start_size,
target_update_interval=1,
minibatch_size=args.batch_size,
target_update_method='soft',
soft_update_tau=5e-3
)
logger = logging.getLogger(__name__)
eval_env = make_env(test=True)
evaluator = Evaluator(
agent=ddpg_agent,
n_steps=None,
n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
max_episode_len=timestep_limit,
env=eval_env,
step_offset=0,
save_best_so_far_agent=True,
use_tensorboard=True,
logger=logger,
)
episode_reward = 0
ddpg_episode_idx = 0
adversary_episode_idx = 0
# o_0, r_0
current_state = env.reset()
t = 0
ddpg_t = 0
adversary_t = 0
episode_len = 0
try:
while t < args.max_steps:
for i in range(args.ddpg_training_steps):
t += 1
ddpg_t += 1
ddpg_action = ddpg_agent.act(current_state)
adversary_action = adversary_agent.act(current_state)
ddpg_action[adversary_action] = 0
next_state, reward, done, info = env.step(ddpg_action)
episode_reward += reward
episode_len += 1
reset = episode_len == timestep_limit or info.get("needs_reset", False)
ddpg_agent.observe(next_state, reward, done, reset)
current_state = next_state
if done or reset or t == args.max_steps:
logger.info(
"ddpg phase: outdir:%s step:%s episode:%s R:%s",
args.outdir,
ddpg_t,
ddpg_episode_idx,
episode_reward,
)
logger.info("statistics:%s", ddpg_agent.get_statistics())
if evaluator is not None:
evaluator.evaluate_if_necessary(t=t, episodes=ddpg_episode_idx + 1)
if t == args.max_steps:
break
episode_reward = 0
ddpg_episode_idx += 1
episode_len = 0
current_state = env.reset()
episode_reward = 0
episode_len = 0
current_state = env.reset()
print("start adversary training ")
for i in range(args.adversary_training_steps):
t += 1
adversary_t += 1
ddpg_action = ddpg_agent.act(current_state)
adversary_action = adversary_agent.act(current_state)
ddpg_action[adversary_action] = 0
next_state, reward, done, info = env.step(ddpg_action)
reward = -reward
episode_len += 1
reset = episode_len == timestep_limit or info.get("needs_reset", False)
adversary_agent.observe(next_state, reward, done, reset)
current_state = next_state
if done or reset or t == args.max_steps:
if t == args.max_steps:
break
episode_reward = 0
adversary_episode_idx += 1
episode_len = 0
current_state = env.reset()
except (Exception, KeyboardInterrupt):
# Save the current model before being killed
save_agent(ddpg_agent, t, args.outdir, logger, suffix="_ddpg_except")
save_agent(adversary_agent, t, args.outdir, logger, suffix="_adversary_except" )
raise
# Save the final model
save_agent(ddpg_agent, t, args.outdir, logger, suffix="_ddpg_finish")
save_agent(adversary_agent, t, args.outdir, logger, suffix="_adversary_finish" )
# if args.demo:
# eval_env.render()
# eval_stats = experiments.eval_performance(
# env=eval_env,
# agent=ddpg_agent,
# n_steps=None,
# n_episodes=args.eval_n_runbase_envs,
# max_episode_len=timestep_limit,
# )
# print(
# "n_runs: {} mean: {} median: {} stdev {}".format(
# args.eval_n_runs,
# eval_stats["mean"],
# eval_stats["median"],
# eval_stats["stdev"],
# )
# )
# else:
# experiments.train_agent_with_evaluation(
# agent=ddpg_agent,
# env=env,
# steps=args.steps,
# eval_env=eval_env,
# eval_n_steps=None,
# eval_n_episodes=args.eval_n_runs,
# eval_interval=args.eval_interval,
# outdir=args.outdir,
# train_max_episode_len=timestep_limit,
# )
print("finish")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--final-exploration-frames", type=int, default=10**6)
parser.add_argument("--final-epsilon", type=float, default=0.01)
parser.add_argument("--eval-epsilon", type=float, default=0.001)
parser.add_argument("--steps", type=int, default=5 * 10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=5 * 10**4)
parser.add_argument("--target-update-interval", type=int, default=10**4)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument("--update-interval", type=int, default=4)
parser.add_argument("--batch-size", type=int, default=32)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--batch-accumulator",
type=str,
default="mean",
choices=["mean", "sum"])
parser.add_argument("--quantile-thresholds-N", type=int, default=64)
parser.add_argument("--quantile-thresholds-N-prime", type=int, default=64)
parser.add_argument("--quantile-thresholds-K", type=int, default=32)
parser.add_argument("--n-best-episodes", type=int, default=200)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = pfrl.agents.iqn.ImplicitQuantileQFunction(
psi=nn.Sequential(
nn.Conv2d(4, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.ReLU(),
nn.Flatten(),
),
phi=nn.Sequential(
pfrl.agents.iqn.CosineBasisLinear(64, 3136),
nn.ReLU(),
),
f=nn.Sequential(
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
),
)
# Use the same hyper parameters as https://arxiv.org/abs/1710.10044
opt = torch.optim.Adam(q_func.parameters(),
lr=5e-5,
eps=1e-2 / args.batch_size)
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.IQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator=args.batch_accumulator,
phi=phi,
quantile_thresholds_N=args.quantile_thresholds_N,
quantile_thresholds_N_prime=args.quantile_thresholds_N_prime,
quantile_thresholds_K=args.quantile_thresholds_K,
)
if args.load or args.load_pretrained:
# either load or load_pretrained must be false
assert not args.load or not args.load_pretrained
if args.load:
agent.load(args.load)
else:
agent.load(
utils.download_model("IQN",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None,
)
print("n_steps: {} mean: {} median: {} stdev {}".format(
args.eval_n_steps,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
dir_of_best_network = os.path.join(args.outdir, "best")
agent.load(dir_of_best_network)
# run 200 evaluation episodes, each capped at 30 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=args.max_frames / 4,
logger=None,
)
with open(os.path.join(args.outdir, "bestscores.json"), "w") as f:
json.dump(stats, f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="OpenAI Atari domain to perform algorithm on.",
)
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--load", type=str, default=None)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument(
"--steps",
type=int,
default=5 * 10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=5 * 10**4,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument("--n-best-episodes", type=int, default=30)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=None),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, 0.05)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = nn.Sequential(
pnn.LargeAtariCNN(),
init_chainer_default(nn.Linear(512, n_actions)),
DiscreteActionValueHead(),
)
# Use the same hyperparameters as the Nature paper
opt = pfrl.optimizers.RMSpropEpsInsideSqrt(
q_func.parameters(),
lr=2.5e-4,
alpha=0.95,
momentum=0.0,
eps=1e-2,
centered=True,
)
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=0.1,
decay_steps=10**6,
random_action_func=lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = agents.DQN
agent = Agent(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=10**4,
clip_delta=True,
update_interval=4,
batch_accumulator="sum",
phi=phi,
)
if args.load or args.load_pretrained:
# either load or load_pretrained must be false
assert not args.load or not args.load_pretrained
if args.load:
agent.load(args.load)
else:
agent.load(
utils.download_model("DQN",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None)
print("n_episodes: {} mean: {} median: {} stdev {}".format(
eval_stats["episodes"],
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
dir_of_best_network = os.path.join(args.outdir, "best")
agent.load(dir_of_best_network)
# run 30 evaluation episodes, each capped at 5 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=4500,
logger=None,
)
with open(os.path.join(args.outdir, "bestscores.json"), "w") as f:
json.dump(stats, f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
def train_PFRL_agent():
policy = make_policy().to(dev)
q_func1 = Q_Net().to(dev)
q_func2 = Q_Net().to(dev)
policy_optimizer = torch.optim.Adam(policy.parameters(), lr=3e-4)
q_func1_optimizer = torch.optim.Adam(q_func1.parameters(), lr=3e-4)
q_func2_optimizer = torch.optim.Adam(q_func2.parameters(), lr=3e-4)
gamma = 0.99
gpu = -1
replay_start_size = 5 * 10 ** 3
minibatch_size = 256
max_grad_norm = 0.5
update_interval = 1
replay_buffer = replay_buffers.ReplayBuffer(5 * 10 ** 3)
def burn_in_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(-1, 1, size=2).astype(np.float32)
print(torch.cuda.is_available())
agent = pfrl.agents.SoftActorCritic(policy,
q_func1, q_func2, policy_optimizer, q_func1_optimizer,
q_func2_optimizer, replay_buffer, gamma, gpu, replay_start_size,
minibatch_size, update_interval, max_grad_norm, temperature_optimizer_lr=3e-4,
burnin_action_func=burn_in_action_func)
env = make_env()
# env1 = make_batch_env(False, env)
# env2 = make_batch_env(True, env)
# experiments.train_agent_batch_with_evaluation(
# agent=agent,
# env=env,
# eval_env=env,
# outdir="./",
# steps=3 * 10 ** 6,
# eval_n_steps=None,
# eval_n_episodes=2,
# eval_interval=2 * 10 ** 3,
# log_interval=10,
# max_episode_len=None,
# )
eval_interval = 2 * 10 ** 1
policy_start_step = 5 * 10 ** 3
state = env.reset()
for i in tqdm(range(3*10**6)):
if i // eval_interval == 0 and i is not 0:
with agent.eval_mode():
state = env.reset()
r_sum = 0
while True:
act = agent.act(state)
n_state, rew, done, info = env.step(act)
r_sum += rew
if done:
print("step {}: rew is {}.".format(i, r_sum))
state = env.reset()
break
act = agent.act(state)
print("act {}".format(act))
n_state, rew, done, info = env.step(act)
agent.observe(n_state, rew, done, done)
if done:
state = env.reset()
def train_PFRL_agent():
env = make_env()
env = pfrl.wrappers.CastObservationToFloat32(env)
env = pfrl.wrappers.NormalizeActionSpace(env)
policy = make_policy(env.state_shape, env.action_space.shape).to(dev)
q_func1 = Q_Net(env.state_shape, env.action_space.shape).to(dev)
q_func2 = Q_Net(env.state_shape, env.action_space.shape).to(dev)
policy_optimizer = torch.optim.Adam(policy.parameters(), lr=3e-4)
q_func1_optimizer = torch.optim.Adam(q_func1.parameters(), lr=3e-4)
q_func2_optimizer = torch.optim.Adam(q_func2.parameters(), lr=3e-4)
gamma = 0.99
gpu = -1
replay_start_size = 5 * 10**1
# replay_start_size = 0
minibatch_size = 1
max_grad_norm = 0.5
update_interval = 1
replay_buffer = replay_buffers.ReplayBuffer(10**6)
def burn_in_action_func():
"""Select random actions until model is updated one or more times."""
print("burn in action func")
ans = np.random.uniform(-1, 1, size=2).astype(np.float32)
print("ans shape {}, ans type {}".format(ans.shape, type(ans)))
return ans
print(torch.cuda.is_available())
agent = pfrl.agents.SoftActorCritic(policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
replay_buffer,
gamma,
gpu,
replay_start_size,
minibatch_size,
update_interval,
max_grad_norm,
temperature_optimizer_lr=3e-4,
burnin_action_func=burn_in_action_func)
eval_interval = 2 * 10**1
policy_start_step = 5 * 10**1
# experiments.train_agent_with_evaluation(
# agent=agent,
# env=env,
# steps=3*10**6,
# eval_n_steps=100,
# eval_n_episodes=None,
# eval_interval=1,
# outdir="./",
# save_best_so_far_agent=True,
# eval_env=env,
# )
state = env.reset()
state_test = state.reshape(1, -1)
print("state shape {}, type {}".format(state.shape, type(state)))
print("state_test shape {}, type {}".format(state_test.shape,
type(state_test)))
with agent.eval_mode():
agent.act(state_test)
for i in tqdm(range(3 * 10**6)):
if i // eval_interval == 0 and i != 0:
with agent.eval_mode():
state = env.reset()
state = torch.from_numpy(state).to(dev)
r_sum = 0
while True:
act = agent.act(state)
n_state, rew, done, info = env.step(act)
r_sum += rew
state = torch.from_numpy(n_state).to(dev)
if done:
print("step {}: rew is {}.".format(i, r_sum))
state = env.reset()
break
# if i < policy_start_step:
# act = env.action_space.sample()
# else:
act = agent.act(state)
print("act {}".format(act))
n_state, rew, done, info = env.step(act)
agent.observe(n_state, rew, done, done)
if done:
state = env.reset()
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument(
"--env",
type=str,
default="Hopper-v2",
help="OpenAI Gym MuJoCo env to perform algorithm on.",
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--load",
type=str,
default="",
help="Directory to load agent from.")
parser.add_argument(
"--steps",
type=int,
default=10**6,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=10,
help="Number of episodes run for each evaluation.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=5000,
help="Interval in timesteps between evaluations.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=10000,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument("--batch-size",
type=int,
default=100,
help="Minibatch size")
parser.add_argument("--render",
action="store_true",
help="Render env states in a GUI window.")
parser.add_argument("--demo",
action="store_true",
help="Just run evaluation, not training.")
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--monitor",
action="store_true",
help="Wrap env with gym.wrappers.Monitor.")
parser.add_argument("--log-level",
type=int,
default=logging.INFO,
help="Level of the root logger.")
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
def make_env(test):
env = gym.make(args.env)
# Unwrap TimeLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.env
# Use different random seeds for train and test envs
env_seed = 2**32 - 1 - args.seed if test else args.seed
env.seed(env_seed)
# Cast observations to float32 because our model uses float32
env = pfrl.wrappers.CastObservationToFloat32(env)
if args.monitor:
env = pfrl.wrappers.Monitor(env, args.outdir)
if args.render and not test:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
timestep_limit = env.spec.max_episode_steps
obs_space = env.observation_space
action_space = env.action_space
print("Observation space:", obs_space)
print("Action space:", action_space)
obs_size = obs_space.low.size
action_size = action_space.low.size
policy = nn.Sequential(
nn.Linear(obs_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, action_size),
nn.Tanh(),
pfrl.policies.DeterministicHead(),
)
policy_optimizer = torch.optim.Adam(policy.parameters())
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 400),
nn.ReLU(),
nn.Linear(400, 300),
nn.ReLU(),
nn.Linear(300, 1),
)
q_func_optimizer = torch.optim.Adam(q_func.parameters())
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.AdditiveGaussian(scale=0.1,
low=action_space.low,
high=action_space.high)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
# Hyperparameters in http://arxiv.org/abs/1802.09477
agent = pfrl.agents.TD3(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=0.99,
soft_update_tau=5e-3,
explorer=explorer,
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=burnin_action_func,
)
if len(args.load) > 0 or args.load_pretrained:
# either load or load_pretrained must be false
assert not len(args.load) > 0 or not args.load_pretrained
if len(args.load) > 0:
agent.load(args.load)
else:
agent.load(
utils.download_model("TD3",
args.env,
model_type=args.pretrained_type)[0])
eval_env = make_env(test=True)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
import json
import os
with open(os.path.join(args.outdir, "demo_scores.json"), "w") as f:
json.dump(eval_stats, f)
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_env=eval_env,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
train_max_episode_len=timestep_limit,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="OpenAI Atari domain to perform algorithm on.",
)
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument(
"--final-exploration-frames",
type=int,
default=10**6,
help="Timesteps after which we stop " + "annealing exploration rate",
)
parser.add_argument(
"--final-epsilon",
type=float,
default=0.01,
help="Final value of epsilon during training.",
)
parser.add_argument(
"--eval-epsilon",
type=float,
default=0.001,
help="Exploration epsilon used during eval episodes.",
)
parser.add_argument("--noisy-net-sigma", type=float, default=None)
parser.add_argument(
"--arch",
type=str,
default="doubledqn",
choices=["nature", "nips", "dueling", "doubledqn"],
help="Network architecture to use.",
)
parser.add_argument(
"--steps",
type=int,
default=5 * 10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=5 * 10**4,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument(
"--target-update-interval",
type=int,
default=3 * 10**4,
help="Frequency (in timesteps) at which " +
"the target network is updated.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=10**5,
help="Frequency (in timesteps) of evaluation phase.",
)
parser.add_argument(
"--update-interval",
type=int,
default=4,
help="Frequency (in timesteps) of network updates.",
)
parser.add_argument("--eval-n-runs", type=int, default=10)
parser.add_argument("--no-clip-delta",
dest="clip_delta",
action="store_false")
parser.add_argument("--num-step-return", type=int, default=1)
parser.set_defaults(clip_delta=True)
parser.add_argument("--agent",
type=str,
default="DoubleDQN",
choices=["DQN", "DoubleDQN", "PAL"])
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--lr",
type=float,
default=2.5e-4,
help="Learning rate.")
parser.add_argument(
"--prioritized",
action="store_true",
default=False,
help="Use prioritized experience replay.",
)
parser.add_argument(
"--checkpoint-frequency",
type=int,
default=None,
help="Frequency at which agents are stored.",
)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = parse_arch(args.arch, n_actions)
if args.noisy_net_sigma is not None:
pnn.to_factorized_noisy(q_func, sigma_scale=args.noisy_net_sigma)
# Turn off explorer
explorer = explorers.Greedy()
else:
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
# Use the Nature paper's hyperparameters
opt = pfrl.optimizers.RMSpropEpsInsideSqrt(
q_func.parameters(),
lr=args.lr,
alpha=0.95,
momentum=0.0,
eps=1e-2,
centered=True,
)
# Select a replay buffer to use
if args.prioritized:
# Anneal beta from beta0 to 1 throughout training
betasteps = args.steps / args.update_interval
rbuf = replay_buffers.PrioritizedReplayBuffer(
10**6,
alpha=0.6,
beta0=0.4,
betasteps=betasteps,
num_steps=args.num_step_return,
)
else:
rbuf = replay_buffers.ReplayBuffer(10**6, args.num_step_return)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
Agent = parse_agent(args.agent)
agent = Agent(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
clip_delta=args.clip_delta,
update_interval=args.update_interval,
batch_accumulator="sum",
phi=phi,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=None,
checkpoint_freq=args.checkpoint_frequency,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=False,
eval_env=eval_env,
)
def main():
if LooseVersion(torch.__version__) < LooseVersion("1.5.0"):
raise Exception("This script requires a PyTorch version >= 1.5.0")
parser = argparse.ArgumentParser()
parser.add_argument('-w',
'--weight_dir',
type=str,
default='',
help='path to trained')
parser.add_argument('-s',
'--step_to_load',
type=int,
default=0,
help='step checkpoint to load')
parser.add_argument('--gpu',
type=int,
default=0,
help='gpu id (-1 for cpu)')
args = parser.parse_args()
weight_dir = args.weight_dir
step_to_load = args.step_to_load
task_path = os.path.dirname(os.path.realpath(__file__))
rsc_path = task_path + "/../rsc"
save_path = os.path.join(
weight_dir, 'testing_' + str(step_to_load),
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
os.makedirs(save_path)
for file in os.listdir(weight_dir):
if file.startswith('cfg_sac'):
cfg_abs_path = weight_dir + '/' + file
# config
cfg = YAML().load(open(cfg_abs_path, 'r'))
cfg['environment']['num_envs'] = 1
cfg['environment']['num_threads'] = 1
cfg['environment']['control_dt'] = cfg['testing']['control_dt']
cfg['environment']['render'] = cfg['testing']['render']
impl = anymal_example_env(rsc_path,
dump(cfg['environment'], Dumper=RoundTripDumper))
env = VecEnvPython(impl)
obs_space = env.observation_space
action_space = env.action_space
print("Observation space:", obs_space)
print("Action space:", action_space)
# seeding
seed = cfg['environment']['seed']
torch.manual_seed(seed)
utils.set_random_seed(seed) # Set a random seed used in PFRL
obs_size = obs_space.low.size
action_size = action_space.low.size
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = torch.chunk(x, 2, dim=1)
log_scale = torch.clamp(log_scale, -20.0, 2.0)
var = torch.exp(log_scale * 2)
base_distribution = distributions.Independent(
distributions.Normal(loc=mean, scale=torch.sqrt(var)), 1)
# cache_size=1 is required for numerical stability
return distributions.transformed_distribution.TransformedDistribution(
base_distribution,
[distributions.transforms.TanhTransform(cache_size=1)])
policy = nn.Sequential(
nn.Linear(obs_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, action_size * 2),
Lambda(squashed_diagonal_gaussian_head),
)
policy_optimizer = torch.optim.Adam(policy.parameters(),
lr=cfg['algorithm']['learning_rate'])
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 1),
)
torch.nn.init.xavier_uniform_(q_func[1].weight)
torch.nn.init.xavier_uniform_(q_func[3].weight)
torch.nn.init.xavier_uniform_(q_func[5].weight)
q_func_optimizer = torch.optim.Adam(
q_func.parameters(), lr=cfg['algorithm']['learning_rate'])
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
rbuf = replay_buffers.ReplayBuffer(cfg['algorithm']['replay_buffer_size'])
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
agent = pfrl.agents.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=cfg['algorithm']['discount_factor'],
replay_start_size=cfg['algorithm']['replay_start_size'],
gpu=args.gpu,
minibatch_size=cfg['algorithm']['minibatch_size'],
burnin_action_func=burnin_action_func,
entropy_target=-action_size,
temperature_optimizer_lr=cfg['algorithm']['temperature_optimizer_lr'],
)
agent.load(weight_dir + '/' + str(step_to_load) + '_checkpoint')
if cfg['testing']['render']:
env.wrapper.showWindow()
if cfg['testing']['record_video']:
env.start_recording_video(save_path + '/test.mp4')
test_steps = int(cfg['testing']['seconds'] / cfg['testing']['control_dt'])
torch.manual_seed(cfg['environment']['seed'])
act = np.ndarray(shape=(1, env.wrapper.getActionDim()), dtype=np.float32)
_, _, _, new_info = env.step(act, visualize=cfg['testing']['render'])
ob = env.reset()
try:
for i in range(test_steps):
if i % 100 == 0:
env.reset()
with agent.eval_mode():
agent.act_deterministically = True
act = agent.batch_act(ob)
ob, rew, done, info = env.step(act,
visualize=cfg['testing']['render'])
except KeyboardInterrupt:
pass
finally:
if cfg['testing']['record_video']:
env.stop_recording_video()
def _test_load_sac(self, gpu):
obs_size = 11
action_size = 3
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = torch.chunk(x, 2, dim=1)
log_scale = torch.clamp(log_scale, -20.0, 2.0)
var = torch.exp(log_scale * 2)
from torch import distributions
base_distribution = distributions.Independent(
distributions.Normal(loc=mean, scale=torch.sqrt(var)), 1)
# cache_size=1 is required for numerical stability
return distributions.transformed_distribution.TransformedDistribution(
base_distribution,
[distributions.transforms.TanhTransform(cache_size=1)],
)
from pfrl.nn.lmbda import Lambda
policy = nn.Sequential(
nn.Linear(obs_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, action_size * 2),
Lambda(squashed_diagonal_gaussian_head),
)
policy_optimizer = torch.optim.Adam(policy.parameters(), lr=3e-4)
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(obs_size + action_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 1),
)
torch.nn.init.xavier_uniform_(q_func[1].weight)
torch.nn.init.xavier_uniform_(q_func[3].weight)
torch.nn.init.xavier_uniform_(q_func[5].weight)
q_func_optimizer = torch.optim.Adam(q_func.parameters(), lr=3e-4)
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
agent = agents.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
replay_buffers.ReplayBuffer(100),
gamma=0.99,
replay_start_size=1000,
gpu=gpu,
minibatch_size=256,
burnin_action_func=None,
entropy_target=-3,
temperature_optimizer_lr=3e-4,
)
downloaded_model, exists = download_model(
"SAC", "Hopper-v2", model_type=self.pretrained_type)
agent.load(downloaded_model)
if os.environ.get("PFRL_ASSERT_DOWNLOADED_MODEL_IS_CACHED"):
assert exists
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--num-envs",
type=int,
default=1,
help="Number of envs run in parallel.")
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 32)")
parser.add_argument("--gpu",
type=int,
default=0,
help="GPU to use, set to -1 if no GPU.")
parser.add_argument("--load",
type=str,
default="",
help="Directory to load agent from.")
parser.add_argument(
"--steps",
type=int,
default=10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--eval-n-runs",
type=int,
default=10,
help="Number of episodes run for each evaluation.",
)
parser.add_argument(
"--eval-interval",
type=int,
default=100000,
help="Interval in timesteps between evaluations.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=2500,
help="Minimum replay buffer size before " +
"performing gradient updates.",
)
parser.add_argument(
"--update-interval",
type=int,
default=1,
help="Interval in timesteps between model updates.",
)
parser.add_argument("--batch-size",
type=int,
default=100,
help="Minibatch size")
parser.add_argument("--render",
action="store_true",
help="Render env states in a GUI window.")
parser.add_argument("--demo",
action="store_true",
help="Just run evaluation, not training.")
parser.add_argument("--monitor",
action="store_true",
help="Wrap env with Monitor to write videos.")
parser.add_argument(
"--log-interval",
type=int,
default=1000,
help=
"Interval in timesteps between outputting log messages during training",
)
parser.add_argument("--log-level",
type=int,
default=logging.INFO,
help="Level of the root logger.")
parser.add_argument(
"--n-hidden-channels",
type=int,
default=256,
help="Number of hidden channels of NN models.",
)
parser.add_argument(
"--env",
default="AntMaze",
help=
"Type of Ant Env to use. Options are AntMaze, AntFall, and AntPush.",
type=str)
parser.add_argument("--discount",
type=float,
default=0.99,
help="Discount factor.")
parser.add_argument("--n-step-return",
type=int,
default=3,
help="N-step return.")
parser.add_argument("--lr",
type=float,
default=3e-4,
help="Learning rate.")
parser.add_argument("--adam-eps",
type=float,
default=1e-1,
help="Adam eps.")
args = parser.parse_args()
logging.basicConfig(level=args.log_level)
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv)
print("Output files are saved in {}".format(args.outdir))
# Set a random seed used in PFRL
utils.set_random_seed(args.seed)
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def make_ant_env(idx, test):
# use different seeds for train vs test envs
process_seed = int(process_seeds[idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
# env_seed = np.random.randint(0, 2**32 - 1) if not test else process_seed
print('seed', env_seed)
utils.set_random_seed(env_seed)
# create the anv environment with goal
env = AntEnvWithGoal(create_maze_env(args.env),
args.env,
env_subgoal_dim=15)
env.seed(int(env_seed))
if args.render:
env = pfrl.wrappers.GymLikeEnvRender(env, mode='human')
return env
eval_env = make_ant_env(0, test=True)
env_state_dim = eval_env.state_dim
env_action_dim = eval_env.action_dim
if args.env == 'AntMaze' or args.env == 'AntPush':
env_goal_dim = 2
else:
env_goal_dim = 3
action_size = env_action_dim
action_space = eval_env.action_space
scale_low = action_space.high * np.ones(env_action_dim)
def squashed_diagonal_gaussian_head(x):
assert x.shape[-1] == action_size * 2
mean, log_scale = torch.chunk(x, 2, dim=1)
log_scale = torch.clamp(log_scale, -20.0, 2.0)
var = torch.exp(log_scale * 2)
base_distribution = distributions.Independent(
distributions.Normal(loc=mean, scale=torch.sqrt(var)), 1)
# cache_size=1 is required for numerical stability
return distributions.transformed_distribution.TransformedDistribution(
base_distribution,
[distributions.transforms.TanhTransform(cache_size=1)])
policy = nn.Sequential(
nn.Linear(env_state_dim + env_goal_dim, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, args.n_hidden_channels),
nn.ReLU(),
nn.Linear(args.n_hidden_channels, action_size * 2),
Lambda(squashed_diagonal_gaussian_head),
)
torch.nn.init.xavier_uniform_(policy[0].weight)
torch.nn.init.xavier_uniform_(policy[2].weight)
torch.nn.init.xavier_uniform_(policy[4].weight)
policy_optimizer = torch.optim.Adam(policy.parameters(), lr=0.0001)
def make_q_func_with_optimizer():
q_func = nn.Sequential(
pfrl.nn.ConcatObsAndAction(),
nn.Linear(env_state_dim + env_goal_dim + env_action_dim, 300),
nn.ReLU(),
nn.Linear(300, 300),
nn.ReLU(),
nn.Linear(300, 1),
)
torch.nn.init.xavier_uniform_(q_func[1].weight)
torch.nn.init.xavier_uniform_(q_func[3].weight)
torch.nn.init.xavier_uniform_(q_func[5].weight)
q_func_optimizer = torch.optim.Adam(q_func.parameters(), lr=0.001)
return q_func, q_func_optimizer
q_func1, q_func1_optimizer = make_q_func_with_optimizer()
q_func2, q_func2_optimizer = make_q_func_with_optimizer()
rbuf = replay_buffers.ReplayBuffer(200000)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
if args.gpu is not None and args.gpu >= 0:
assert torch.cuda.is_available()
device = torch.device("cuda:{}".format(args.gpu))
else:
device = torch.device("cpu")
# Hyperparameters in http://arxiv.org/abs/1802.09477
scale_tensor = torch.tensor(scale_low).float().to(device)
agent = pfrl.agents.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=args.discount,
update_interval=args.update_interval,
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
burnin_action_func=burnin_action_func,
entropy_target=-action_size,
temperature_optimizer_lr=args.lr,
scale=scale_tensor)
if len(args.load) > 0:
agent.load(args.load)
if args.demo:
eval_env = make_env(args, seed=0, test=True)
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print("n_runs: {} mean: {} median: {} stdev {}".format(
args.eval_n_runs,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_goal_conditioned_agent_with_evaluation(
agent=agent,
env=make_ant_env(0, test=False),
steps=args.steps,
eval_n_steps=None,
outdir=args.outdir,
eval_n_episodes=args.eval_n_runs,
eval_interval=5000,
use_tensorboard=True,
)